arch/x86/kvm/svm/sev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel-based Virtual Machine driver for Linux
   4  *
   5  * AMD SVM-SEV support
   6  *
   7  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   8  */
   9
  10 #include <linux/kvm_types.h>
  11 #include <linux/kvm_host.h>
  12 #include <linux/kernel.h>
  13 #include <linux/highmem.h>
  14 #include <linux/psp-sev.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/swap.h>
  17 #include <linux/misc_cgroup.h>
  18 #include <linux/processor.h>
  19 #include <linux/trace_events.h>
  20
  21 #include <asm/pkru.h>
  22 #include <asm/trapnr.h>
  23 #include <asm/fpu/xcr.h>
  24
  25 #include "x86.h"
  26 #include "svm.h"
  27 #include "svm_ops.h"
  28 #include "cpuid.h"
  29 #include "trace.h"
  30
  31 #ifndef CONFIG_KVM_AMD_SEV
  32 /*
  33  * When this config is not defined, SEV feature is not supported and APIs in
  34  * this file are not used but this file still gets compiled into the KVM AMD
  35  * module.
  36  *
  37  * We will not have MISC_CG_RES_SEV and MISC_CG_RES_SEV_ES entries in the enum
  38  * misc_res_type {} defined in linux/misc_cgroup.h.
  39  *
  40  * Below macros allow compilation to succeed.
  41  */
  42 #define MISC_CG_RES_SEV MISC_CG_RES_TYPES
  43 #define MISC_CG_RES_SEV_ES MISC_CG_RES_TYPES
  44 #endif
  45
  46 #ifdef CONFIG_KVM_AMD_SEV
  47 /* enable/disable SEV support */
  48 static bool sev_enabled = true;
  49 module_param_named(sev, sev_enabled, bool, 0444);
  50
  51 /* enable/disable SEV-ES support */
  52 static bool sev_es_enabled = true;
  53 module_param_named(sev_es, sev_es_enabled, bool, 0444);
  54 #else
  55 #define sev_enabled false
  56 #define sev_es_enabled false
  57 #endif /* CONFIG_KVM_AMD_SEV */
  58
  59 static u8 sev_enc_bit;
  60 static DECLARE_RWSEM(sev_deactivate_lock);
  61 static DEFINE_MUTEX(sev_bitmap_lock);
  62 unsigned int max_sev_asid;
  63 static unsigned int min_sev_asid;
  64 static unsigned long sev_me_mask;
  65 static unsigned int nr_asids;
  66 static unsigned long *sev_asid_bitmap;
  67 static unsigned long *sev_reclaim_asid_bitmap;
  68
  69 struct enc_region {
  70         struct list_head list;
  71         unsigned long npages;
  72         struct page **pages;
  73         unsigned long uaddr;
  74         unsigned long size;
  75 };
  76
  77 /* Called with the sev_bitmap_lock held, or on shutdown  */
  78 static int sev_flush_asids(int min_asid, int max_asid)
  79 {
  80         int ret, asid, error = 0;
  81
  82         /* Check if there are any ASIDs to reclaim before performing a flush */
  83         asid = find_next_bit(sev_reclaim_asid_bitmap, nr_asids, min_asid);
  84         if (asid > max_asid)
  85                 return -EBUSY;
  86
  87         /*
  88          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  89          * so it must be guarded.
  90          */
  91         down_write(&sev_deactivate_lock);
  92
  93         wbinvd_on_all_cpus();
  94         ret = sev_guest_df_flush(&error);
  95
  96         up_write(&sev_deactivate_lock);
  97
  98         if (ret)
  99                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
 100
 101         return ret;
 102 }
 103
 104 static inline bool is_mirroring_enc_context(struct kvm *kvm)
 105 {
 106         return !!to_kvm_svm(kvm)->sev_info.enc_context_owner;
 107 }
 108
 109 /* Must be called with the sev_bitmap_lock held */
 110 static bool __sev_recycle_asids(int min_asid, int max_asid)
 111 {
 112         if (sev_flush_asids(min_asid, max_asid))
 113                 return false;
 114
 115         /* The flush process will flush all reclaimable SEV and SEV-ES ASIDs */
 116         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
 117                    nr_asids);
 118         bitmap_zero(sev_reclaim_asid_bitmap, nr_asids);
 119
 120         return true;
 121 }
 122
 123 static int sev_misc_cg_try_charge(struct kvm_sev_info *sev)
 124 {
 125         enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 126         return misc_cg_try_charge(type, sev->misc_cg, 1);
 127 }
 128
 129 static void sev_misc_cg_uncharge(struct kvm_sev_info *sev)
 130 {
 131         enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 132         misc_cg_uncharge(type, sev->misc_cg, 1);
 133 }
 134
 135 static int sev_asid_new(struct kvm_sev_info *sev)
 136 {
 137         int asid, min_asid, max_asid, ret;
 138         bool retry = true;
 139
 140         WARN_ON(sev->misc_cg);
 141         sev->misc_cg = get_current_misc_cg();
 142         ret = sev_misc_cg_try_charge(sev);
 143         if (ret) {
 144                 put_misc_cg(sev->misc_cg);
 145                 sev->misc_cg = NULL;
 146                 return ret;
 147         }
 148
 149         mutex_lock(&sev_bitmap_lock);
 150
 151         /*
 152          * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
 153          * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
 154          */
 155         min_asid = sev->es_active ? 1 : min_sev_asid;
 156         max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
 157 again:
 158         asid = find_next_zero_bit(sev_asid_bitmap, max_asid + 1, min_asid);
 159         if (asid > max_asid) {
 160                 if (retry && __sev_recycle_asids(min_asid, max_asid)) {
 161                         retry = false;
 162                         goto again;
 163                 }
 164                 mutex_unlock(&sev_bitmap_lock);
 165                 ret = -EBUSY;
 166                 goto e_uncharge;
 167         }
 168
 169         __set_bit(asid, sev_asid_bitmap);
 170
 171         mutex_unlock(&sev_bitmap_lock);
 172
 173         return asid;
 174 e_uncharge:
 175         sev_misc_cg_uncharge(sev);
 176         put_misc_cg(sev->misc_cg);
 177         sev->misc_cg = NULL;
 178         return ret;
 179 }
 180
 181 static int sev_get_asid(struct kvm *kvm)
 182 {
 183         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 184
 185         return sev->asid;
 186 }
 187
 188 static void sev_asid_free(struct kvm_sev_info *sev)
 189 {
 190         struct svm_cpu_data *sd;
 191         int cpu;
 192
 193         mutex_lock(&sev_bitmap_lock);
 194
 195         __set_bit(sev->asid, sev_reclaim_asid_bitmap);
 196
 197         for_each_possible_cpu(cpu) {
 198                 sd = per_cpu(svm_data, cpu);
 199                 sd->sev_vmcbs[sev->asid] = NULL;
 200         }
 201
 202         mutex_unlock(&sev_bitmap_lock);
 203
 204         sev_misc_cg_uncharge(sev);
 205         put_misc_cg(sev->misc_cg);
 206         sev->misc_cg = NULL;
 207 }
 208
 209 static void sev_decommission(unsigned int handle)
 210 {
 211         struct sev_data_decommission decommission;
 212
 213         if (!handle)
 214                 return;
 215
 216         decommission.handle = handle;
 217         sev_guest_decommission(&decommission, NULL);
 218 }
 219
 220 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 221 {
 222         struct sev_data_deactivate deactivate;
 223
 224         if (!handle)
 225                 return;
 226
 227         deactivate.handle = handle;
 228
 229         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 230         down_read(&sev_deactivate_lock);
 231         sev_guest_deactivate(&deactivate, NULL);
 232         up_read(&sev_deactivate_lock);
 233
 234         sev_decommission(handle);
 235 }
 236
 237 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 238 {
 239         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 240         int asid, ret;
 241
 242         if (kvm->created_vcpus)
 243                 return -EINVAL;
 244
 245         ret = -EBUSY;
 246         if (unlikely(sev->active))
 247                 return ret;
 248
 249         sev->active = true;
 250         sev->es_active = argp->id == KVM_SEV_ES_INIT;
 251         asid = sev_asid_new(sev);
 252         if (asid < 0)
 253                 goto e_no_asid;
 254         sev->asid = asid;
 255
 256         ret = sev_platform_init(&argp->error);
 257         if (ret)
 258                 goto e_free;
 259
 260         INIT_LIST_HEAD(&sev->regions_list);
 261         INIT_LIST_HEAD(&sev->mirror_vms);
 262
 263         kvm_set_apicv_inhibit(kvm, APICV_INHIBIT_REASON_SEV);
 264
 265         return 0;
 266
 267 e_free:
 268         sev_asid_free(sev);
 269         sev->asid = 0;
 270 e_no_asid:
 271         sev->es_active = false;
 272         sev->active = false;
 273         return ret;
 274 }
 275
 276 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 277 {
 278         struct sev_data_activate activate;
 279         int asid = sev_get_asid(kvm);
 280         int ret;
 281
 282         /* activate ASID on the given handle */
 283         activate.handle = handle;
 284         activate.asid   = asid;
 285         ret = sev_guest_activate(&activate, error);
 286
 287         return ret;
 288 }
 289
 290 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 291 {
 292         struct fd f;
 293         int ret;
 294
 295         f = fdget(fd);
 296         if (!f.file)
 297                 return -EBADF;
 298
 299         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 300
 301         fdput(f);
 302         return ret;
 303 }
 304
 305 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 306 {
 307         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 308
 309         return __sev_issue_cmd(sev->fd, id, data, error);
 310 }
 311
 312 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 313 {
 314         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 315         struct sev_data_launch_start start;
 316         struct kvm_sev_launch_start params;
 317         void *dh_blob, *session_blob;
 318         int *error = &argp->error;
 319         int ret;
 320
 321         if (!sev_guest(kvm))
 322                 return -ENOTTY;
 323
 324         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 325                 return -EFAULT;
 326
 327         memset(&start, 0, sizeof(start));
 328
 329         dh_blob = NULL;
 330         if (params.dh_uaddr) {
 331                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 332                 if (IS_ERR(dh_blob))
 333                         return PTR_ERR(dh_blob);
 334
 335                 start.dh_cert_address = __sme_set(__pa(dh_blob));
 336                 start.dh_cert_len = params.dh_len;
 337         }
 338
 339         session_blob = NULL;
 340         if (params.session_uaddr) {
 341                 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
 342                 if (IS_ERR(session_blob)) {
 343                         ret = PTR_ERR(session_blob);
 344                         goto e_free_dh;
 345                 }
 346
 347                 start.session_address = __sme_set(__pa(session_blob));
 348                 start.session_len = params.session_len;
 349         }
 350
 351         start.handle = params.handle;
 352         start.policy = params.policy;
 353
 354         /* create memory encryption context */
 355         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, &start, error);
 356         if (ret)
 357                 goto e_free_session;
 358
 359         /* Bind ASID to this guest */
 360         ret = sev_bind_asid(kvm, start.handle, error);
 361         if (ret) {
 362                 sev_decommission(start.handle);
 363                 goto e_free_session;
 364         }
 365
 366         /* return handle to userspace */
 367         params.handle = start.handle;
 368         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 369                 sev_unbind_asid(kvm, start.handle);
 370                 ret = -EFAULT;
 371                 goto e_free_session;
 372         }
 373
 374         sev->handle = start.handle;
 375         sev->fd = argp->sev_fd;
 376
 377 e_free_session:
 378         kfree(session_blob);
 379 e_free_dh:
 380         kfree(dh_blob);
 381         return ret;
 382 }
 383
 384 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
 385                                     unsigned long ulen, unsigned long *n,
 386                                     int write)
 387 {
 388         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 389         unsigned long npages, size;
 390         int npinned;
 391         unsigned long locked, lock_limit;
 392         struct page **pages;
 393         unsigned long first, last;
 394         int ret;
 395
 396         lockdep_assert_held(&kvm->lock);
 397
 398         if (ulen == 0 || uaddr + ulen < uaddr)
 399                 return ERR_PTR(-EINVAL);
 400
 401         /* Calculate number of pages. */
 402         first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
 403         last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
 404         npages = (last - first + 1);
 405
 406         locked = sev->pages_locked + npages;
 407         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 408         if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
 409                 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
 410                 return ERR_PTR(-ENOMEM);
 411         }
 412
 413         if (WARN_ON_ONCE(npages > INT_MAX))
 414                 return ERR_PTR(-EINVAL);
 415
 416         /* Avoid using vmalloc for smaller buffers. */
 417         size = npages * sizeof(struct page *);
 418         if (size > PAGE_SIZE)
 419                 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 420         else
 421                 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
 422
 423         if (!pages)
 424                 return ERR_PTR(-ENOMEM);
 425
 426         /* Pin the user virtual address. */
 427         npinned = pin_user_pages_fast(uaddr, npages, write ? FOLL_WRITE : 0, pages);
 428         if (npinned != npages) {
 429                 pr_err("SEV: Failure locking %lu pages.\n", npages);
 430                 ret = -ENOMEM;
 431                 goto err;
 432         }
 433
 434         *n = npages;
 435         sev->pages_locked = locked;
 436
 437         return pages;
 438
 439 err:
 440         if (npinned > 0)
 441                 unpin_user_pages(pages, npinned);
 442
 443         kvfree(pages);
 444         return ERR_PTR(ret);
 445 }
 446
 447 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
 448                              unsigned long npages)
 449 {
 450         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 451
 452         unpin_user_pages(pages, npages);
 453         kvfree(pages);
 454         sev->pages_locked -= npages;
 455 }
 456
 457 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
 458 {
 459         uint8_t *page_virtual;
 460         unsigned long i;
 461
 462         if (this_cpu_has(X86_FEATURE_SME_COHERENT) || npages == 0 ||
 463             pages == NULL)
 464                 return;
 465
 466         for (i = 0; i < npages; i++) {
 467                 page_virtual = kmap_atomic(pages[i]);
 468                 clflush_cache_range(page_virtual, PAGE_SIZE);
 469                 kunmap_atomic(page_virtual);
 470                 cond_resched();
 471         }
 472 }
 473
 474 static unsigned long get_num_contig_pages(unsigned long idx,
 475                                 struct page **inpages, unsigned long npages)
 476 {
 477         unsigned long paddr, next_paddr;
 478         unsigned long i = idx + 1, pages = 1;
 479
 480         /* find the number of contiguous pages starting from idx */
 481         paddr = __sme_page_pa(inpages[idx]);
 482         while (i < npages) {
 483                 next_paddr = __sme_page_pa(inpages[i++]);
 484                 if ((paddr + PAGE_SIZE) == next_paddr) {
 485                         pages++;
 486                         paddr = next_paddr;
 487                         continue;
 488                 }
 489                 break;
 490         }
 491
 492         return pages;
 493 }
 494
 495 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
 496 {
 497         unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
 498         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 499         struct kvm_sev_launch_update_data params;
 500         struct sev_data_launch_update_data data;
 501         struct page **inpages;
 502         int ret;
 503
 504         if (!sev_guest(kvm))
 505                 return -ENOTTY;
 506
 507         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 508                 return -EFAULT;
 509
 510         vaddr = params.uaddr;
 511         size = params.len;
 512         vaddr_end = vaddr + size;
 513
 514         /* Lock the user memory. */
 515         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 516         if (IS_ERR(inpages))
 517                 return PTR_ERR(inpages);
 518
 519         /*
 520          * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
 521          * place; the cache may contain the data that was written unencrypted.
 522          */
 523         sev_clflush_pages(inpages, npages);
 524
 525         data.reserved = 0;
 526         data.handle = sev->handle;
 527
 528         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 529                 int offset, len;
 530
 531                 /*
 532                  * If the user buffer is not page-aligned, calculate the offset
 533                  * within the page.
 534                  */
 535                 offset = vaddr & (PAGE_SIZE - 1);
 536
 537                 /* Calculate the number of pages that can be encrypted in one go. */
 538                 pages = get_num_contig_pages(i, inpages, npages);
 539
 540                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 541
 542                 data.len = len;
 543                 data.address = __sme_page_pa(inpages[i]) + offset;
 544                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, &data, &argp->error);
 545                 if (ret)
 546                         goto e_unpin;
 547
 548                 size -= len;
 549                 next_vaddr = vaddr + len;
 550         }
 551
 552 e_unpin:
 553         /* content of memory is updated, mark pages dirty */
 554         for (i = 0; i < npages; i++) {
 555                 set_page_dirty_lock(inpages[i]);
 556                 mark_page_accessed(inpages[i]);
 557         }
 558         /* unlock the user pages */
 559         sev_unpin_memory(kvm, inpages, npages);
 560         return ret;
 561 }
 562
 563 static int sev_es_sync_vmsa(struct vcpu_svm *svm)
 564 {
 565         struct vmcb_save_area *save = &svm->vmcb->save;
 566
 567         /* Check some debug related fields before encrypting the VMSA */
 568         if (svm->vcpu.guest_debug || (save->dr7 & ~DR7_FIXED_1))
 569                 return -EINVAL;
 570
 571         /* Sync registgers */
 572         save->rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
 573         save->rbx = svm->vcpu.arch.regs[VCPU_REGS_RBX];
 574         save->rcx = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 575         save->rdx = svm->vcpu.arch.regs[VCPU_REGS_RDX];
 576         save->rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
 577         save->rbp = svm->vcpu.arch.regs[VCPU_REGS_RBP];
 578         save->rsi = svm->vcpu.arch.regs[VCPU_REGS_RSI];
 579         save->rdi = svm->vcpu.arch.regs[VCPU_REGS_RDI];
 580 #ifdef CONFIG_X86_64
 581         save->r8  = svm->vcpu.arch.regs[VCPU_REGS_R8];
 582         save->r9  = svm->vcpu.arch.regs[VCPU_REGS_R9];
 583         save->r10 = svm->vcpu.arch.regs[VCPU_REGS_R10];
 584         save->r11 = svm->vcpu.arch.regs[VCPU_REGS_R11];
 585         save->r12 = svm->vcpu.arch.regs[VCPU_REGS_R12];
 586         save->r13 = svm->vcpu.arch.regs[VCPU_REGS_R13];
 587         save->r14 = svm->vcpu.arch.regs[VCPU_REGS_R14];
 588         save->r15 = svm->vcpu.arch.regs[VCPU_REGS_R15];
 589 #endif
 590         save->rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
 591
 592         /* Sync some non-GPR registers before encrypting */
 593         save->xcr0 = svm->vcpu.arch.xcr0;
 594         save->pkru = svm->vcpu.arch.pkru;
 595         save->xss  = svm->vcpu.arch.ia32_xss;
 596         save->dr6  = svm->vcpu.arch.dr6;
 597
 598         /*
 599          * SEV-ES will use a VMSA that is pointed to by the VMCB, not
 600          * the traditional VMSA that is part of the VMCB. Copy the
 601          * traditional VMSA as it has been built so far (in prep
 602          * for LAUNCH_UPDATE_VMSA) to be the initial SEV-ES state.
 603          */
 604         memcpy(svm->sev_es.vmsa, save, sizeof(*save));
 605
 606         return 0;
 607 }
 608
 609 static int __sev_launch_update_vmsa(struct kvm *kvm, struct kvm_vcpu *vcpu,
 610                                     int *error)
 611 {
 612         struct sev_data_launch_update_vmsa vmsa;
 613         struct vcpu_svm *svm = to_svm(vcpu);
 614         int ret;
 615
 616         /* Perform some pre-encryption checks against the VMSA */
 617         ret = sev_es_sync_vmsa(svm);
 618         if (ret)
 619                 return ret;
 620
 621         /*
 622          * The LAUNCH_UPDATE_VMSA command will perform in-place encryption of
 623          * the VMSA memory content (i.e it will write the same memory region
 624          * with the guest's key), so invalidate it first.
 625          */
 626         clflush_cache_range(svm->sev_es.vmsa, PAGE_SIZE);
 627
 628         vmsa.reserved = 0;
 629         vmsa.handle = to_kvm_svm(kvm)->sev_info.handle;
 630         vmsa.address = __sme_pa(svm->sev_es.vmsa);
 631         vmsa.len = PAGE_SIZE;
 632         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_VMSA, &vmsa, error);
 633         if (ret)
 634           return ret;
 635
 636         vcpu->arch.guest_state_protected = true;
 637         return 0;
 638 }
 639
 640 static int sev_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
 641 {
 642         struct kvm_vcpu *vcpu;
 643         unsigned long i;
 644         int ret;
 645
 646         if (!sev_es_guest(kvm))
 647                 return -ENOTTY;
 648
 649         kvm_for_each_vcpu(i, vcpu, kvm) {
 650                 ret = mutex_lock_killable(&vcpu->mutex);
 651                 if (ret)
 652                         return ret;
 653
 654                 ret = __sev_launch_update_vmsa(kvm, vcpu, &argp->error);
 655
 656                 mutex_unlock(&vcpu->mutex);
 657                 if (ret)
 658                         return ret;
 659         }
 660
 661         return 0;
 662 }
 663
 664 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 665 {
 666         void __user *measure = (void __user *)(uintptr_t)argp->data;
 667         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 668         struct sev_data_launch_measure data;
 669         struct kvm_sev_launch_measure params;
 670         void __user *p = NULL;
 671         void *blob = NULL;
 672         int ret;
 673
 674         if (!sev_guest(kvm))
 675                 return -ENOTTY;
 676
 677         if (copy_from_user(&params, measure, sizeof(params)))
 678                 return -EFAULT;
 679
 680         memset(&data, 0, sizeof(data));
 681
 682         /* User wants to query the blob length */
 683         if (!params.len)
 684                 goto cmd;
 685
 686         p = (void __user *)(uintptr_t)params.uaddr;
 687         if (p) {
 688                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
 689                         return -EINVAL;
 690
 691                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
 692                 if (!blob)
 693                         return -ENOMEM;
 694
 695                 data.address = __psp_pa(blob);
 696                 data.len = params.len;
 697         }
 698
 699 cmd:
 700         data.handle = sev->handle;
 701         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, &data, &argp->error);
 702
 703         /*
 704          * If we query the session length, FW responded with expected data.
 705          */
 706         if (!params.len)
 707                 goto done;
 708
 709         if (ret)
 710                 goto e_free_blob;
 711
 712         if (blob) {
 713                 if (copy_to_user(p, blob, params.len))
 714                         ret = -EFAULT;
 715         }
 716
 717 done:
 718         params.len = data.len;
 719         if (copy_to_user(measure, &params, sizeof(params)))
 720                 ret = -EFAULT;
 721 e_free_blob:
 722         kfree(blob);
 723         return ret;
 724 }
 725
 726 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 727 {
 728         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 729         struct sev_data_launch_finish data;
 730
 731         if (!sev_guest(kvm))
 732                 return -ENOTTY;
 733
 734         data.handle = sev->handle;
 735         return sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, &data, &argp->error);
 736 }
 737
 738 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 739 {
 740         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 741         struct kvm_sev_guest_status params;
 742         struct sev_data_guest_status data;
 743         int ret;
 744
 745         if (!sev_guest(kvm))
 746                 return -ENOTTY;
 747
 748         memset(&data, 0, sizeof(data));
 749
 750         data.handle = sev->handle;
 751         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, &data, &argp->error);
 752         if (ret)
 753                 return ret;
 754
 755         params.policy = data.policy;
 756         params.state = data.state;
 757         params.handle = data.handle;
 758
 759         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 760                 ret = -EFAULT;
 761
 762         return ret;
 763 }
 764
 765 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 766                                unsigned long dst, int size,
 767                                int *error, bool enc)
 768 {
 769         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 770         struct sev_data_dbg data;
 771
 772         data.reserved = 0;
 773         data.handle = sev->handle;
 774         data.dst_addr = dst;
 775         data.src_addr = src;
 776         data.len = size;
 777
 778         return sev_issue_cmd(kvm,
 779                              enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 780                              &data, error);
 781 }
 782
 783 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 784                              unsigned long dst_paddr, int sz, int *err)
 785 {
 786         int offset;
 787
 788         /*
 789          * Its safe to read more than we are asked, caller should ensure that
 790          * destination has enough space.
 791          */
 792         offset = src_paddr & 15;
 793         src_paddr = round_down(src_paddr, 16);
 794         sz = round_up(sz + offset, 16);
 795
 796         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 797 }
 798
 799 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 800                                   void __user *dst_uaddr,
 801                                   unsigned long dst_paddr,
 802                                   int size, int *err)
 803 {
 804         struct page *tpage = NULL;
 805         int ret, offset;
 806
 807         /* if inputs are not 16-byte then use intermediate buffer */
 808         if (!IS_ALIGNED(dst_paddr, 16) ||
 809             !IS_ALIGNED(paddr,     16) ||
 810             !IS_ALIGNED(size,      16)) {
 811                 tpage = (void *)alloc_page(GFP_KERNEL);
 812                 if (!tpage)
 813                         return -ENOMEM;
 814
 815                 dst_paddr = __sme_page_pa(tpage);
 816         }
 817
 818         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 819         if (ret)
 820                 goto e_free;
 821
 822         if (tpage) {
 823                 offset = paddr & 15;
 824                 if (copy_to_user(dst_uaddr, page_address(tpage) + offset, size))
 825                         ret = -EFAULT;
 826         }
 827
 828 e_free:
 829         if (tpage)
 830                 __free_page(tpage);
 831
 832         return ret;
 833 }
 834
 835 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 836                                   void __user *vaddr,
 837                                   unsigned long dst_paddr,
 838                                   void __user *dst_vaddr,
 839                                   int size, int *error)
 840 {
 841         struct page *src_tpage = NULL;
 842         struct page *dst_tpage = NULL;
 843         int ret, len = size;
 844
 845         /* If source buffer is not aligned then use an intermediate buffer */
 846         if (!IS_ALIGNED((unsigned long)vaddr, 16)) {
 847                 src_tpage = alloc_page(GFP_KERNEL);
 848                 if (!src_tpage)
 849                         return -ENOMEM;
 850
 851                 if (copy_from_user(page_address(src_tpage), vaddr, size)) {
 852                         __free_page(src_tpage);
 853                         return -EFAULT;
 854                 }
 855
 856                 paddr = __sme_page_pa(src_tpage);
 857         }
 858
 859         /*
 860          *  If destination buffer or length is not aligned then do read-modify-write:
 861          *   - decrypt destination in an intermediate buffer
 862          *   - copy the source buffer in an intermediate buffer
 863          *   - use the intermediate buffer as source buffer
 864          */
 865         if (!IS_ALIGNED((unsigned long)dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 866                 int dst_offset;
 867
 868                 dst_tpage = alloc_page(GFP_KERNEL);
 869                 if (!dst_tpage) {
 870                         ret = -ENOMEM;
 871                         goto e_free;
 872                 }
 873
 874                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 875                                         __sme_page_pa(dst_tpage), size, error);
 876                 if (ret)
 877                         goto e_free;
 878
 879                 /*
 880                  *  If source is kernel buffer then use memcpy() otherwise
 881                  *  copy_from_user().
 882                  */
 883                 dst_offset = dst_paddr & 15;
 884
 885                 if (src_tpage)
 886                         memcpy(page_address(dst_tpage) + dst_offset,
 887                                page_address(src_tpage), size);
 888                 else {
 889                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 890                                            vaddr, size)) {
 891                                 ret = -EFAULT;
 892                                 goto e_free;
 893                         }
 894                 }
 895
 896                 paddr = __sme_page_pa(dst_tpage);
 897                 dst_paddr = round_down(dst_paddr, 16);
 898                 len = round_up(size, 16);
 899         }
 900
 901         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 902
 903 e_free:
 904         if (src_tpage)
 905                 __free_page(src_tpage);
 906         if (dst_tpage)
 907                 __free_page(dst_tpage);
 908         return ret;
 909 }
 910
 911 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 912 {
 913         unsigned long vaddr, vaddr_end, next_vaddr;
 914         unsigned long dst_vaddr;
 915         struct page **src_p, **dst_p;
 916         struct kvm_sev_dbg debug;
 917         unsigned long n;
 918         unsigned int size;
 919         int ret;
 920
 921         if (!sev_guest(kvm))
 922                 return -ENOTTY;
 923
 924         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 925                 return -EFAULT;
 926
 927         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 928                 return -EINVAL;
 929         if (!debug.dst_uaddr)
 930                 return -EINVAL;
 931
 932         vaddr = debug.src_uaddr;
 933         size = debug.len;
 934         vaddr_end = vaddr + size;
 935         dst_vaddr = debug.dst_uaddr;
 936
 937         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 938                 int len, s_off, d_off;
 939
 940                 /* lock userspace source and destination page */
 941                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 942                 if (IS_ERR(src_p))
 943                         return PTR_ERR(src_p);
 944
 945                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 946                 if (IS_ERR(dst_p)) {
 947                         sev_unpin_memory(kvm, src_p, n);
 948                         return PTR_ERR(dst_p);
 949                 }
 950
 951                 /*
 952                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 953                  * the pages; flush the destination too so that future accesses do not
 954                  * see stale data.
 955                  */
 956                 sev_clflush_pages(src_p, 1);
 957                 sev_clflush_pages(dst_p, 1);
 958
 959                 /*
 960                  * Since user buffer may not be page aligned, calculate the
 961                  * offset within the page.
 962                  */
 963                 s_off = vaddr & ~PAGE_MASK;
 964                 d_off = dst_vaddr & ~PAGE_MASK;
 965                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 966
 967                 if (dec)
 968                         ret = __sev_dbg_decrypt_user(kvm,
 969                                                      __sme_page_pa(src_p[0]) + s_off,
 970                                                      (void __user *)dst_vaddr,
 971                                                      __sme_page_pa(dst_p[0]) + d_off,
 972                                                      len, &argp->error);
 973                 else
 974                         ret = __sev_dbg_encrypt_user(kvm,
 975                                                      __sme_page_pa(src_p[0]) + s_off,
 976                                                      (void __user *)vaddr,
 977                                                      __sme_page_pa(dst_p[0]) + d_off,
 978                                                      (void __user *)dst_vaddr,
 979                                                      len, &argp->error);
 980
 981                 sev_unpin_memory(kvm, src_p, n);
 982                 sev_unpin_memory(kvm, dst_p, n);
 983
 984                 if (ret)
 985                         goto err;
 986
 987                 next_vaddr = vaddr + len;
 988                 dst_vaddr = dst_vaddr + len;
 989                 size -= len;
 990         }
 991 err:
 992         return ret;
 993 }
 994
 995 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
 996 {
 997         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 998         struct sev_data_launch_secret data;
 999         struct kvm_sev_launch_secret params;
1000         struct page **pages;
1001         void *blob, *hdr;
1002         unsigned long n, i;
1003         int ret, offset;
1004
1005         if (!sev_guest(kvm))
1006                 return -ENOTTY;
1007
1008         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1009                 return -EFAULT;
1010
1011         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
1012         if (IS_ERR(pages))
1013                 return PTR_ERR(pages);
1014
1015         /*
1016          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
1017          * place; the cache may contain the data that was written unencrypted.
1018          */
1019         sev_clflush_pages(pages, n);
1020
1021         /*
1022          * The secret must be copied into contiguous memory region, lets verify
1023          * that userspace memory pages are contiguous before we issue command.
1024          */
1025         if (get_num_contig_pages(0, pages, n) != n) {
1026                 ret = -EINVAL;
1027                 goto e_unpin_memory;
1028         }
1029
1030         memset(&data, 0, sizeof(data));
1031
1032         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1033         data.guest_address = __sme_page_pa(pages[0]) + offset;
1034         data.guest_len = params.guest_len;
1035
1036         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1037         if (IS_ERR(blob)) {
1038                 ret = PTR_ERR(blob);
1039                 goto e_unpin_memory;
1040         }
1041
1042         data.trans_address = __psp_pa(blob);
1043         data.trans_len = params.trans_len;
1044
1045         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1046         if (IS_ERR(hdr)) {
1047                 ret = PTR_ERR(hdr);
1048                 goto e_free_blob;
1049         }
1050         data.hdr_address = __psp_pa(hdr);
1051         data.hdr_len = params.hdr_len;
1052
1053         data.handle = sev->handle;
1054         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, &data, &argp->error);
1055
1056         kfree(hdr);
1057
1058 e_free_blob:
1059         kfree(blob);
1060 e_unpin_memory:
1061         /* content of memory is updated, mark pages dirty */
1062         for (i = 0; i < n; i++) {
1063                 set_page_dirty_lock(pages[i]);
1064                 mark_page_accessed(pages[i]);
1065         }
1066         sev_unpin_memory(kvm, pages, n);
1067         return ret;
1068 }
1069
1070 static int sev_get_attestation_report(struct kvm *kvm, struct kvm_sev_cmd *argp)
1071 {
1072         void __user *report = (void __user *)(uintptr_t)argp->data;
1073         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1074         struct sev_data_attestation_report data;
1075         struct kvm_sev_attestation_report params;
1076         void __user *p;
1077         void *blob = NULL;
1078         int ret;
1079
1080         if (!sev_guest(kvm))
1081                 return -ENOTTY;
1082
1083         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1084                 return -EFAULT;
1085
1086         memset(&data, 0, sizeof(data));
1087
1088         /* User wants to query the blob length */
1089         if (!params.len)
1090                 goto cmd;
1091
1092         p = (void __user *)(uintptr_t)params.uaddr;
1093         if (p) {
1094                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
1095                         return -EINVAL;
1096
1097                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
1098                 if (!blob)
1099                         return -ENOMEM;
1100
1101                 data.address = __psp_pa(blob);
1102                 data.len = params.len;
1103                 memcpy(data.mnonce, params.mnonce, sizeof(params.mnonce));
1104         }
1105 cmd:
1106         data.handle = sev->handle;
1107         ret = sev_issue_cmd(kvm, SEV_CMD_ATTESTATION_REPORT, &data, &argp->error);
1108         /*
1109          * If we query the session length, FW responded with expected data.
1110          */
1111         if (!params.len)
1112                 goto done;
1113
1114         if (ret)
1115                 goto e_free_blob;
1116
1117         if (blob) {
1118                 if (copy_to_user(p, blob, params.len))
1119                         ret = -EFAULT;
1120         }
1121
1122 done:
1123         params.len = data.len;
1124         if (copy_to_user(report, &params, sizeof(params)))
1125                 ret = -EFAULT;
1126 e_free_blob:
1127         kfree(blob);
1128         return ret;
1129 }
1130
1131 /* Userspace wants to query session length. */
1132 static int
1133 __sev_send_start_query_session_length(struct kvm *kvm, struct kvm_sev_cmd *argp,
1134                                       struct kvm_sev_send_start *params)
1135 {
1136         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1137         struct sev_data_send_start data;
1138         int ret;
1139
1140         memset(&data, 0, sizeof(data));
1141         data.handle = sev->handle;
1142         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1143
1144         params->session_len = data.session_len;
1145         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1146                                 sizeof(struct kvm_sev_send_start)))
1147                 ret = -EFAULT;
1148
1149         return ret;
1150 }
1151
1152 static int sev_send_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1153 {
1154         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1155         struct sev_data_send_start data;
1156         struct kvm_sev_send_start params;
1157         void *amd_certs, *session_data;
1158         void *pdh_cert, *plat_certs;
1159         int ret;
1160
1161         if (!sev_guest(kvm))
1162                 return -ENOTTY;
1163
1164         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1165                                 sizeof(struct kvm_sev_send_start)))
1166                 return -EFAULT;
1167
1168         /* if session_len is zero, userspace wants to query the session length */
1169         if (!params.session_len)
1170                 return __sev_send_start_query_session_length(kvm, argp,
1171                                 &params);
1172
1173         /* some sanity checks */
1174         if (!params.pdh_cert_uaddr || !params.pdh_cert_len ||
1175             !params.session_uaddr || params.session_len > SEV_FW_BLOB_MAX_SIZE)
1176                 return -EINVAL;
1177
1178         /* allocate the memory to hold the session data blob */
1179         session_data = kmalloc(params.session_len, GFP_KERNEL_ACCOUNT);
1180         if (!session_data)
1181                 return -ENOMEM;
1182
1183         /* copy the certificate blobs from userspace */
1184         pdh_cert = psp_copy_user_blob(params.pdh_cert_uaddr,
1185                                 params.pdh_cert_len);
1186         if (IS_ERR(pdh_cert)) {
1187                 ret = PTR_ERR(pdh_cert);
1188                 goto e_free_session;
1189         }
1190
1191         plat_certs = psp_copy_user_blob(params.plat_certs_uaddr,
1192                                 params.plat_certs_len);
1193         if (IS_ERR(plat_certs)) {
1194                 ret = PTR_ERR(plat_certs);
1195                 goto e_free_pdh;
1196         }
1197
1198         amd_certs = psp_copy_user_blob(params.amd_certs_uaddr,
1199                                 params.amd_certs_len);
1200         if (IS_ERR(amd_certs)) {
1201                 ret = PTR_ERR(amd_certs);
1202                 goto e_free_plat_cert;
1203         }
1204
1205         /* populate the FW SEND_START field with system physical address */
1206         memset(&data, 0, sizeof(data));
1207         data.pdh_cert_address = __psp_pa(pdh_cert);
1208         data.pdh_cert_len = params.pdh_cert_len;
1209         data.plat_certs_address = __psp_pa(plat_certs);
1210         data.plat_certs_len = params.plat_certs_len;
1211         data.amd_certs_address = __psp_pa(amd_certs);
1212         data.amd_certs_len = params.amd_certs_len;
1213         data.session_address = __psp_pa(session_data);
1214         data.session_len = params.session_len;
1215         data.handle = sev->handle;
1216
1217         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1218
1219         if (!ret && copy_to_user((void __user *)(uintptr_t)params.session_uaddr,
1220                         session_data, params.session_len)) {
1221                 ret = -EFAULT;
1222                 goto e_free_amd_cert;
1223         }
1224
1225         params.policy = data.policy;
1226         params.session_len = data.session_len;
1227         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params,
1228                                 sizeof(struct kvm_sev_send_start)))
1229                 ret = -EFAULT;
1230
1231 e_free_amd_cert:
1232         kfree(amd_certs);
1233 e_free_plat_cert:
1234         kfree(plat_certs);
1235 e_free_pdh:
1236         kfree(pdh_cert);
1237 e_free_session:
1238         kfree(session_data);
1239         return ret;
1240 }
1241
1242 /* Userspace wants to query either header or trans length. */
1243 static int
1244 __sev_send_update_data_query_lengths(struct kvm *kvm, struct kvm_sev_cmd *argp,
1245                                      struct kvm_sev_send_update_data *params)
1246 {
1247         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1248         struct sev_data_send_update_data data;
1249         int ret;
1250
1251         memset(&data, 0, sizeof(data));
1252         data.handle = sev->handle;
1253         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1254
1255         params->hdr_len = data.hdr_len;
1256         params->trans_len = data.trans_len;
1257
1258         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1259                          sizeof(struct kvm_sev_send_update_data)))
1260                 ret = -EFAULT;
1261
1262         return ret;
1263 }
1264
1265 static int sev_send_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1266 {
1267         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1268         struct sev_data_send_update_data data;
1269         struct kvm_sev_send_update_data params;
1270         void *hdr, *trans_data;
1271         struct page **guest_page;
1272         unsigned long n;
1273         int ret, offset;
1274
1275         if (!sev_guest(kvm))
1276                 return -ENOTTY;
1277
1278         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1279                         sizeof(struct kvm_sev_send_update_data)))
1280                 return -EFAULT;
1281
1282         /* userspace wants to query either header or trans length */
1283         if (!params.trans_len || !params.hdr_len)
1284                 return __sev_send_update_data_query_lengths(kvm, argp, &params);
1285
1286         if (!params.trans_uaddr || !params.guest_uaddr ||
1287             !params.guest_len || !params.hdr_uaddr)
1288                 return -EINVAL;
1289
1290         /* Check if we are crossing the page boundary */
1291         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1292         if ((params.guest_len + offset > PAGE_SIZE))
1293                 return -EINVAL;
1294
1295         /* Pin guest memory */
1296         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1297                                     PAGE_SIZE, &n, 0);
1298         if (IS_ERR(guest_page))
1299                 return PTR_ERR(guest_page);
1300
1301         /* allocate memory for header and transport buffer */
1302         ret = -ENOMEM;
1303         hdr = kmalloc(params.hdr_len, GFP_KERNEL_ACCOUNT);
1304         if (!hdr)
1305                 goto e_unpin;
1306
1307         trans_data = kmalloc(params.trans_len, GFP_KERNEL_ACCOUNT);
1308         if (!trans_data)
1309                 goto e_free_hdr;
1310
1311         memset(&data, 0, sizeof(data));
1312         data.hdr_address = __psp_pa(hdr);
1313         data.hdr_len = params.hdr_len;
1314         data.trans_address = __psp_pa(trans_data);
1315         data.trans_len = params.trans_len;
1316
1317         /* The SEND_UPDATE_DATA command requires C-bit to be always set. */
1318         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1319         data.guest_address |= sev_me_mask;
1320         data.guest_len = params.guest_len;
1321         data.handle = sev->handle;
1322
1323         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1324
1325         if (ret)
1326                 goto e_free_trans_data;
1327
1328         /* copy transport buffer to user space */
1329         if (copy_to_user((void __user *)(uintptr_t)params.trans_uaddr,
1330                          trans_data, params.trans_len)) {
1331                 ret = -EFAULT;
1332                 goto e_free_trans_data;
1333         }
1334
1335         /* Copy packet header to userspace. */
1336         if (copy_to_user((void __user *)(uintptr_t)params.hdr_uaddr, hdr,
1337                          params.hdr_len))
1338                 ret = -EFAULT;
1339
1340 e_free_trans_data:
1341         kfree(trans_data);
1342 e_free_hdr:
1343         kfree(hdr);
1344 e_unpin:
1345         sev_unpin_memory(kvm, guest_page, n);
1346
1347         return ret;
1348 }
1349
1350 static int sev_send_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1351 {
1352         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1353         struct sev_data_send_finish data;
1354
1355         if (!sev_guest(kvm))
1356                 return -ENOTTY;
1357
1358         data.handle = sev->handle;
1359         return sev_issue_cmd(kvm, SEV_CMD_SEND_FINISH, &data, &argp->error);
1360 }
1361
1362 static int sev_send_cancel(struct kvm *kvm, struct kvm_sev_cmd *argp)
1363 {
1364         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1365         struct sev_data_send_cancel data;
1366
1367         if (!sev_guest(kvm))
1368                 return -ENOTTY;
1369
1370         data.handle = sev->handle;
1371         return sev_issue_cmd(kvm, SEV_CMD_SEND_CANCEL, &data, &argp->error);
1372 }
1373
1374 static int sev_receive_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1375 {
1376         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1377         struct sev_data_receive_start start;
1378         struct kvm_sev_receive_start params;
1379         int *error = &argp->error;
1380         void *session_data;
1381         void *pdh_data;
1382         int ret;
1383
1384         if (!sev_guest(kvm))
1385                 return -ENOTTY;
1386
1387         /* Get parameter from the userspace */
1388         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1389                         sizeof(struct kvm_sev_receive_start)))
1390                 return -EFAULT;
1391
1392         /* some sanity checks */
1393         if (!params.pdh_uaddr || !params.pdh_len ||
1394             !params.session_uaddr || !params.session_len)
1395                 return -EINVAL;
1396
1397         pdh_data = psp_copy_user_blob(params.pdh_uaddr, params.pdh_len);
1398         if (IS_ERR(pdh_data))
1399                 return PTR_ERR(pdh_data);
1400
1401         session_data = psp_copy_user_blob(params.session_uaddr,
1402                         params.session_len);
1403         if (IS_ERR(session_data)) {
1404                 ret = PTR_ERR(session_data);
1405                 goto e_free_pdh;
1406         }
1407
1408         memset(&start, 0, sizeof(start));
1409         start.handle = params.handle;
1410         start.policy = params.policy;
1411         start.pdh_cert_address = __psp_pa(pdh_data);
1412         start.pdh_cert_len = params.pdh_len;
1413         start.session_address = __psp_pa(session_data);
1414         start.session_len = params.session_len;
1415
1416         /* create memory encryption context */
1417         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_RECEIVE_START, &start,
1418                                 error);
1419         if (ret)
1420                 goto e_free_session;
1421
1422         /* Bind ASID to this guest */
1423         ret = sev_bind_asid(kvm, start.handle, error);
1424         if (ret) {
1425                 sev_decommission(start.handle);
1426                 goto e_free_session;
1427         }
1428
1429         params.handle = start.handle;
1430         if (copy_to_user((void __user *)(uintptr_t)argp->data,
1431                          &params, sizeof(struct kvm_sev_receive_start))) {
1432                 ret = -EFAULT;
1433                 sev_unbind_asid(kvm, start.handle);
1434                 goto e_free_session;
1435         }
1436
1437         sev->handle = start.handle;
1438         sev->fd = argp->sev_fd;
1439
1440 e_free_session:
1441         kfree(session_data);
1442 e_free_pdh:
1443         kfree(pdh_data);
1444
1445         return ret;
1446 }
1447
1448 static int sev_receive_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1449 {
1450         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1451         struct kvm_sev_receive_update_data params;
1452         struct sev_data_receive_update_data data;
1453         void *hdr = NULL, *trans = NULL;
1454         struct page **guest_page;
1455         unsigned long n;
1456         int ret, offset;
1457
1458         if (!sev_guest(kvm))
1459                 return -EINVAL;
1460
1461         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1462                         sizeof(struct kvm_sev_receive_update_data)))
1463                 return -EFAULT;
1464
1465         if (!params.hdr_uaddr || !params.hdr_len ||
1466             !params.guest_uaddr || !params.guest_len ||
1467             !params.trans_uaddr || !params.trans_len)
1468                 return -EINVAL;
1469
1470         /* Check if we are crossing the page boundary */
1471         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1472         if ((params.guest_len + offset > PAGE_SIZE))
1473                 return -EINVAL;
1474
1475         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1476         if (IS_ERR(hdr))
1477                 return PTR_ERR(hdr);
1478
1479         trans = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1480         if (IS_ERR(trans)) {
1481                 ret = PTR_ERR(trans);
1482                 goto e_free_hdr;
1483         }
1484
1485         memset(&data, 0, sizeof(data));
1486         data.hdr_address = __psp_pa(hdr);
1487         data.hdr_len = params.hdr_len;
1488         data.trans_address = __psp_pa(trans);
1489         data.trans_len = params.trans_len;
1490
1491         /* Pin guest memory */
1492         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1493                                     PAGE_SIZE, &n, 1);
1494         if (IS_ERR(guest_page)) {
1495                 ret = PTR_ERR(guest_page);
1496                 goto e_free_trans;
1497         }
1498
1499         /*
1500          * Flush (on non-coherent CPUs) before RECEIVE_UPDATE_DATA, the PSP
1501          * encrypts the written data with the guest's key, and the cache may
1502          * contain dirty, unencrypted data.
1503          */
1504         sev_clflush_pages(guest_page, n);
1505
1506         /* The RECEIVE_UPDATE_DATA command requires C-bit to be always set. */
1507         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1508         data.guest_address |= sev_me_mask;
1509         data.guest_len = params.guest_len;
1510         data.handle = sev->handle;
1511
1512         ret = sev_issue_cmd(kvm, SEV_CMD_RECEIVE_UPDATE_DATA, &data,
1513                                 &argp->error);
1514
1515         sev_unpin_memory(kvm, guest_page, n);
1516
1517 e_free_trans:
1518         kfree(trans);
1519 e_free_hdr:
1520         kfree(hdr);
1521
1522         return ret;
1523 }
1524
1525 static int sev_receive_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1526 {
1527         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1528         struct sev_data_receive_finish data;
1529
1530         if (!sev_guest(kvm))
1531                 return -ENOTTY;
1532
1533         data.handle = sev->handle;
1534         return sev_issue_cmd(kvm, SEV_CMD_RECEIVE_FINISH, &data, &argp->error);
1535 }
1536
1537 static bool is_cmd_allowed_from_mirror(u32 cmd_id)
1538 {
1539         /*
1540          * Allow mirrors VM to call KVM_SEV_LAUNCH_UPDATE_VMSA to enable SEV-ES
1541          * active mirror VMs. Also allow the debugging and status commands.
1542          */
1543         if (cmd_id == KVM_SEV_LAUNCH_UPDATE_VMSA ||
1544             cmd_id == KVM_SEV_GUEST_STATUS || cmd_id == KVM_SEV_DBG_DECRYPT ||
1545             cmd_id == KVM_SEV_DBG_ENCRYPT)
1546                 return true;
1547
1548         return false;
1549 }
1550
1551 static int sev_lock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1552 {
1553         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1554         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1555         int r = -EBUSY;
1556
1557         if (dst_kvm == src_kvm)
1558                 return -EINVAL;
1559
1560         /*
1561          * Bail if these VMs are already involved in a migration to avoid
1562          * deadlock between two VMs trying to migrate to/from each other.
1563          */
1564         if (atomic_cmpxchg_acquire(&dst_sev->migration_in_progress, 0, 1))
1565                 return -EBUSY;
1566
1567         if (atomic_cmpxchg_acquire(&src_sev->migration_in_progress, 0, 1))
1568                 goto release_dst;
1569
1570         r = -EINTR;
1571         if (mutex_lock_killable(&dst_kvm->lock))
1572                 goto release_src;
1573         if (mutex_lock_killable_nested(&src_kvm->lock, SINGLE_DEPTH_NESTING))
1574                 goto unlock_dst;
1575         return 0;
1576
1577 unlock_dst:
1578         mutex_unlock(&dst_kvm->lock);
1579 release_src:
1580         atomic_set_release(&src_sev->migration_in_progress, 0);
1581 release_dst:
1582         atomic_set_release(&dst_sev->migration_in_progress, 0);
1583         return r;
1584 }
1585
1586 static void sev_unlock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1587 {
1588         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1589         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1590
1591         mutex_unlock(&dst_kvm->lock);
1592         mutex_unlock(&src_kvm->lock);
1593         atomic_set_release(&dst_sev->migration_in_progress, 0);
1594         atomic_set_release(&src_sev->migration_in_progress, 0);
1595 }
1596
1597 /* vCPU mutex subclasses.  */
1598 enum sev_migration_role {
1599         SEV_MIGRATION_SOURCE = 0,
1600         SEV_MIGRATION_TARGET,
1601         SEV_NR_MIGRATION_ROLES,
1602 };
1603
1604 static int sev_lock_vcpus_for_migration(struct kvm *kvm,
1605                                         enum sev_migration_role role)
1606 {
1607         struct kvm_vcpu *vcpu;
1608         unsigned long i, j;
1609         bool first = true;
1610
1611         kvm_for_each_vcpu(i, vcpu, kvm) {
1612                 if (mutex_lock_killable_nested(&vcpu->mutex, role))
1613                         goto out_unlock;
1614
1615                 if (first) {
1616                         /*
1617                          * Reset the role to one that avoids colliding with
1618                          * the role used for the first vcpu mutex.
1619                          */
1620                         role = SEV_NR_MIGRATION_ROLES;
1621                         first = false;
1622                 } else {
1623                         mutex_release(&vcpu->mutex.dep_map, _THIS_IP_);
1624                 }
1625         }
1626
1627         return 0;
1628
1629 out_unlock:
1630
1631         first = true;
1632         kvm_for_each_vcpu(j, vcpu, kvm) {
1633                 if (i == j)
1634                         break;
1635
1636                 if (first)
1637                         first = false;
1638                 else
1639                         mutex_acquire(&vcpu->mutex.dep_map, role, 0, _THIS_IP_);
1640
1641
1642                 mutex_unlock(&vcpu->mutex);
1643         }
1644         return -EINTR;
1645 }
1646
1647 static void sev_unlock_vcpus_for_migration(struct kvm *kvm)
1648 {
1649         struct kvm_vcpu *vcpu;
1650         unsigned long i;
1651         bool first = true;
1652
1653         kvm_for_each_vcpu(i, vcpu, kvm) {
1654                 if (first)
1655                         first = false;
1656                 else
1657                         mutex_acquire(&vcpu->mutex.dep_map,
1658                                       SEV_NR_MIGRATION_ROLES, 0, _THIS_IP_);
1659
1660                 mutex_unlock(&vcpu->mutex);
1661         }
1662 }
1663
1664 static void sev_migrate_from(struct kvm *dst_kvm, struct kvm *src_kvm)
1665 {
1666         struct kvm_sev_info *dst = &to_kvm_svm(dst_kvm)->sev_info;
1667         struct kvm_sev_info *src = &to_kvm_svm(src_kvm)->sev_info;
1668         struct kvm_sev_info *mirror;
1669
1670         dst->active = true;
1671         dst->asid = src->asid;
1672         dst->handle = src->handle;
1673         dst->pages_locked = src->pages_locked;
1674         dst->enc_context_owner = src->enc_context_owner;
1675
1676         src->asid = 0;
1677         src->active = false;
1678         src->handle = 0;
1679         src->pages_locked = 0;
1680         src->enc_context_owner = NULL;
1681
1682         list_cut_before(&dst->regions_list, &src->regions_list, &src->regions_list);
1683
1684         /*
1685          * If this VM has mirrors, "transfer" each mirror's refcount of the
1686          * source to the destination (this KVM).  The caller holds a reference
1687          * to the source, so there's no danger of use-after-free.
1688          */
1689         list_cut_before(&dst->mirror_vms, &src->mirror_vms, &src->mirror_vms);
1690         list_for_each_entry(mirror, &dst->mirror_vms, mirror_entry) {
1691                 kvm_get_kvm(dst_kvm);
1692                 kvm_put_kvm(src_kvm);
1693                 mirror->enc_context_owner = dst_kvm;
1694         }
1695
1696         /*
1697          * If this VM is a mirror, remove the old mirror from the owners list
1698          * and add the new mirror to the list.
1699          */
1700         if (is_mirroring_enc_context(dst_kvm)) {
1701                 struct kvm_sev_info *owner_sev_info =
1702                         &to_kvm_svm(dst->enc_context_owner)->sev_info;
1703
1704                 list_del(&src->mirror_entry);
1705                 list_add_tail(&dst->mirror_entry, &owner_sev_info->mirror_vms);
1706         }
1707 }
1708
1709 static int sev_es_migrate_from(struct kvm *dst, struct kvm *src)
1710 {
1711         unsigned long i;
1712         struct kvm_vcpu *dst_vcpu, *src_vcpu;
1713         struct vcpu_svm *dst_svm, *src_svm;
1714
1715         if (atomic_read(&src->online_vcpus) != atomic_read(&dst->online_vcpus))
1716                 return -EINVAL;
1717
1718         kvm_for_each_vcpu(i, src_vcpu, src) {
1719                 if (!src_vcpu->arch.guest_state_protected)
1720                         return -EINVAL;
1721         }
1722
1723         kvm_for_each_vcpu(i, src_vcpu, src) {
1724                 src_svm = to_svm(src_vcpu);
1725                 dst_vcpu = kvm_get_vcpu(dst, i);
1726                 dst_svm = to_svm(dst_vcpu);
1727
1728                 /*
1729                  * Transfer VMSA and GHCB state to the destination.  Nullify and
1730                  * clear source fields as appropriate, the state now belongs to
1731                  * the destination.
1732                  */
1733                 memcpy(&dst_svm->sev_es, &src_svm->sev_es, sizeof(src_svm->sev_es));
1734                 dst_svm->vmcb->control.ghcb_gpa = src_svm->vmcb->control.ghcb_gpa;
1735                 dst_svm->vmcb->control.vmsa_pa = src_svm->vmcb->control.vmsa_pa;
1736                 dst_vcpu->arch.guest_state_protected = true;
1737
1738                 memset(&src_svm->sev_es, 0, sizeof(src_svm->sev_es));
1739                 src_svm->vmcb->control.ghcb_gpa = INVALID_PAGE;
1740                 src_svm->vmcb->control.vmsa_pa = INVALID_PAGE;
1741                 src_vcpu->arch.guest_state_protected = false;
1742         }
1743         to_kvm_svm(src)->sev_info.es_active = false;
1744         to_kvm_svm(dst)->sev_info.es_active = true;
1745
1746         return 0;
1747 }
1748
1749 int sev_vm_move_enc_context_from(struct kvm *kvm, unsigned int source_fd)
1750 {
1751         struct kvm_sev_info *dst_sev = &to_kvm_svm(kvm)->sev_info;
1752         struct kvm_sev_info *src_sev, *cg_cleanup_sev;
1753         struct file *source_kvm_file;
1754         struct kvm *source_kvm;
1755         bool charged = false;
1756         int ret;
1757
1758         source_kvm_file = fget(source_fd);
1759         if (!file_is_kvm(source_kvm_file)) {
1760                 ret = -EBADF;
1761                 goto out_fput;
1762         }
1763
1764         source_kvm = source_kvm_file->private_data;
1765         ret = sev_lock_two_vms(kvm, source_kvm);
1766         if (ret)
1767                 goto out_fput;
1768
1769         if (sev_guest(kvm) || !sev_guest(source_kvm)) {
1770                 ret = -EINVAL;
1771                 goto out_unlock;
1772         }
1773
1774         src_sev = &to_kvm_svm(source_kvm)->sev_info;
1775
1776         dst_sev->misc_cg = get_current_misc_cg();
1777         cg_cleanup_sev = dst_sev;
1778         if (dst_sev->misc_cg != src_sev->misc_cg) {
1779                 ret = sev_misc_cg_try_charge(dst_sev);
1780                 if (ret)
1781                         goto out_dst_cgroup;
1782                 charged = true;
1783         }
1784
1785         ret = sev_lock_vcpus_for_migration(kvm, SEV_MIGRATION_SOURCE);
1786         if (ret)
1787                 goto out_dst_cgroup;
1788         ret = sev_lock_vcpus_for_migration(source_kvm, SEV_MIGRATION_TARGET);
1789         if (ret)
1790                 goto out_dst_vcpu;
1791
1792         if (sev_es_guest(source_kvm)) {
1793                 ret = sev_es_migrate_from(kvm, source_kvm);
1794                 if (ret)
1795                         goto out_source_vcpu;
1796         }
1797
1798         sev_migrate_from(kvm, source_kvm);
1799         kvm_vm_dead(source_kvm);
1800         cg_cleanup_sev = src_sev;
1801         ret = 0;
1802
1803 out_source_vcpu:
1804         sev_unlock_vcpus_for_migration(source_kvm);
1805 out_dst_vcpu:
1806         sev_unlock_vcpus_for_migration(kvm);
1807 out_dst_cgroup:
1808         /* Operates on the source on success, on the destination on failure.  */
1809         if (charged)
1810                 sev_misc_cg_uncharge(cg_cleanup_sev);
1811         put_misc_cg(cg_cleanup_sev->misc_cg);
1812         cg_cleanup_sev->misc_cg = NULL;
1813 out_unlock:
1814         sev_unlock_two_vms(kvm, source_kvm);
1815 out_fput:
1816         if (source_kvm_file)
1817                 fput(source_kvm_file);
1818         return ret;
1819 }
1820
1821 int sev_mem_enc_ioctl(struct kvm *kvm, void __user *argp)
1822 {
1823         struct kvm_sev_cmd sev_cmd;
1824         int r;
1825
1826         if (!sev_enabled)
1827                 return -ENOTTY;
1828
1829         if (!argp)
1830                 return 0;
1831
1832         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
1833                 return -EFAULT;
1834
1835         mutex_lock(&kvm->lock);
1836
1837         /* Only the enc_context_owner handles some memory enc operations. */
1838         if (is_mirroring_enc_context(kvm) &&
1839             !is_cmd_allowed_from_mirror(sev_cmd.id)) {
1840                 r = -EINVAL;
1841                 goto out;
1842         }
1843
1844         switch (sev_cmd.id) {
1845         case KVM_SEV_ES_INIT:
1846                 if (!sev_es_enabled) {
1847                         r = -ENOTTY;
1848                         goto out;
1849                 }
1850                 fallthrough;
1851         case KVM_SEV_INIT:
1852                 r = sev_guest_init(kvm, &sev_cmd);
1853                 break;
1854         case KVM_SEV_LAUNCH_START:
1855                 r = sev_launch_start(kvm, &sev_cmd);
1856                 break;
1857         case KVM_SEV_LAUNCH_UPDATE_DATA:
1858                 r = sev_launch_update_data(kvm, &sev_cmd);
1859                 break;
1860         case KVM_SEV_LAUNCH_UPDATE_VMSA:
1861                 r = sev_launch_update_vmsa(kvm, &sev_cmd);
1862                 break;
1863         case KVM_SEV_LAUNCH_MEASURE:
1864                 r = sev_launch_measure(kvm, &sev_cmd);
1865                 break;
1866         case KVM_SEV_LAUNCH_FINISH:
1867                 r = sev_launch_finish(kvm, &sev_cmd);
1868                 break;
1869         case KVM_SEV_GUEST_STATUS:
1870                 r = sev_guest_status(kvm, &sev_cmd);
1871                 break;
1872         case KVM_SEV_DBG_DECRYPT:
1873                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
1874                 break;
1875         case KVM_SEV_DBG_ENCRYPT:
1876                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
1877                 break;
1878         case KVM_SEV_LAUNCH_SECRET:
1879                 r = sev_launch_secret(kvm, &sev_cmd);
1880                 break;
1881         case KVM_SEV_GET_ATTESTATION_REPORT:
1882                 r = sev_get_attestation_report(kvm, &sev_cmd);
1883                 break;
1884         case KVM_SEV_SEND_START:
1885                 r = sev_send_start(kvm, &sev_cmd);
1886                 break;
1887         case KVM_SEV_SEND_UPDATE_DATA:
1888                 r = sev_send_update_data(kvm, &sev_cmd);
1889                 break;
1890         case KVM_SEV_SEND_FINISH:
1891                 r = sev_send_finish(kvm, &sev_cmd);
1892                 break;
1893         case KVM_SEV_SEND_CANCEL:
1894                 r = sev_send_cancel(kvm, &sev_cmd);
1895                 break;
1896         case KVM_SEV_RECEIVE_START:
1897                 r = sev_receive_start(kvm, &sev_cmd);
1898                 break;
1899         case KVM_SEV_RECEIVE_UPDATE_DATA:
1900                 r = sev_receive_update_data(kvm, &sev_cmd);
1901                 break;
1902         case KVM_SEV_RECEIVE_FINISH:
1903                 r = sev_receive_finish(kvm, &sev_cmd);
1904                 break;
1905         default:
1906                 r = -EINVAL;
1907                 goto out;
1908         }
1909
1910         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
1911                 r = -EFAULT;
1912
1913 out:
1914         mutex_unlock(&kvm->lock);
1915         return r;
1916 }
1917
1918 int sev_mem_enc_register_region(struct kvm *kvm,
1919                                 struct kvm_enc_region *range)
1920 {
1921         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1922         struct enc_region *region;
1923         int ret = 0;
1924
1925         if (!sev_guest(kvm))
1926                 return -ENOTTY;
1927
1928         /* If kvm is mirroring encryption context it isn't responsible for it */
1929         if (is_mirroring_enc_context(kvm))
1930                 return -EINVAL;
1931
1932         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1933                 return -EINVAL;
1934
1935         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1936         if (!region)
1937                 return -ENOMEM;
1938
1939         mutex_lock(&kvm->lock);
1940         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1941         if (IS_ERR(region->pages)) {
1942                 ret = PTR_ERR(region->pages);
1943                 mutex_unlock(&kvm->lock);
1944                 goto e_free;
1945         }
1946
1947         region->uaddr = range->addr;
1948         region->size = range->size;
1949
1950         list_add_tail(&region->list, &sev->regions_list);
1951         mutex_unlock(&kvm->lock);
1952
1953         /*
1954          * The guest may change the memory encryption attribute from C=0 -> C=1
1955          * or vice versa for this memory range. Lets make sure caches are
1956          * flushed to ensure that guest data gets written into memory with
1957          * correct C-bit.
1958          */
1959         sev_clflush_pages(region->pages, region->npages);
1960
1961         return ret;
1962
1963 e_free:
1964         kfree(region);
1965         return ret;
1966 }
1967
1968 static struct enc_region *
1969 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1970 {
1971         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1972         struct list_head *head = &sev->regions_list;
1973         struct enc_region *i;
1974
1975         list_for_each_entry(i, head, list) {
1976                 if (i->uaddr == range->addr &&
1977                     i->size == range->size)
1978                         return i;
1979         }
1980
1981         return NULL;
1982 }
1983
1984 static void __unregister_enc_region_locked(struct kvm *kvm,
1985                                            struct enc_region *region)
1986 {
1987         sev_unpin_memory(kvm, region->pages, region->npages);
1988         list_del(&region->list);
1989         kfree(region);
1990 }
1991
1992 int sev_mem_enc_unregister_region(struct kvm *kvm,
1993                                   struct kvm_enc_region *range)
1994 {
1995         struct enc_region *region;
1996         int ret;
1997
1998         /* If kvm is mirroring encryption context it isn't responsible for it */
1999         if (is_mirroring_enc_context(kvm))
2000                 return -EINVAL;
2001
2002         mutex_lock(&kvm->lock);
2003
2004         if (!sev_guest(kvm)) {
2005                 ret = -ENOTTY;
2006                 goto failed;
2007         }
2008
2009         region = find_enc_region(kvm, range);
2010         if (!region) {
2011                 ret = -EINVAL;
2012                 goto failed;
2013         }
2014
2015         /*
2016          * Ensure that all guest tagged cache entries are flushed before
2017          * releasing the pages back to the system for use. CLFLUSH will
2018          * not do this, so issue a WBINVD.
2019          */
2020         wbinvd_on_all_cpus();
2021
2022         __unregister_enc_region_locked(kvm, region);
2023
2024         mutex_unlock(&kvm->lock);
2025         return 0;
2026
2027 failed:
2028         mutex_unlock(&kvm->lock);
2029         return ret;
2030 }
2031
2032 int sev_vm_copy_enc_context_from(struct kvm *kvm, unsigned int source_fd)
2033 {
2034         struct file *source_kvm_file;
2035         struct kvm *source_kvm;
2036         struct kvm_sev_info *source_sev, *mirror_sev;
2037         int ret;
2038
2039         source_kvm_file = fget(source_fd);
2040         if (!file_is_kvm(source_kvm_file)) {
2041                 ret = -EBADF;
2042                 goto e_source_fput;
2043         }
2044
2045         source_kvm = source_kvm_file->private_data;
2046         ret = sev_lock_two_vms(kvm, source_kvm);
2047         if (ret)
2048                 goto e_source_fput;
2049
2050         /*
2051          * Mirrors of mirrors should work, but let's not get silly.  Also
2052          * disallow out-of-band SEV/SEV-ES init if the target is already an
2053          * SEV guest, or if vCPUs have been created.  KVM relies on vCPUs being
2054          * created after SEV/SEV-ES initialization, e.g. to init intercepts.
2055          */
2056         if (sev_guest(kvm) || !sev_guest(source_kvm) ||
2057             is_mirroring_enc_context(source_kvm) || kvm->created_vcpus) {
2058                 ret = -EINVAL;
2059                 goto e_unlock;
2060         }
2061
2062         /*
2063          * The mirror kvm holds an enc_context_owner ref so its asid can't
2064          * disappear until we're done with it
2065          */
2066         source_sev = &to_kvm_svm(source_kvm)->sev_info;
2067         kvm_get_kvm(source_kvm);
2068         mirror_sev = &to_kvm_svm(kvm)->sev_info;
2069         list_add_tail(&mirror_sev->mirror_entry, &source_sev->mirror_vms);
2070
2071         /* Set enc_context_owner and copy its encryption context over */
2072         mirror_sev->enc_context_owner = source_kvm;
2073         mirror_sev->active = true;
2074         mirror_sev->asid = source_sev->asid;
2075         mirror_sev->fd = source_sev->fd;
2076         mirror_sev->es_active = source_sev->es_active;
2077         mirror_sev->handle = source_sev->handle;
2078         INIT_LIST_HEAD(&mirror_sev->regions_list);
2079         INIT_LIST_HEAD(&mirror_sev->mirror_vms);
2080         ret = 0;
2081
2082         /*
2083          * Do not copy ap_jump_table. Since the mirror does not share the same
2084          * KVM contexts as the original, and they may have different
2085          * memory-views.
2086          */
2087
2088 e_unlock:
2089         sev_unlock_two_vms(kvm, source_kvm);
2090 e_source_fput:
2091         if (source_kvm_file)
2092                 fput(source_kvm_file);
2093         return ret;
2094 }
2095
2096 void sev_vm_destroy(struct kvm *kvm)
2097 {
2098         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
2099         struct list_head *head = &sev->regions_list;
2100         struct list_head *pos, *q;
2101
2102         if (!sev_guest(kvm))
2103                 return;
2104
2105         WARN_ON(!list_empty(&sev->mirror_vms));
2106
2107         /* If this is a mirror_kvm release the enc_context_owner and skip sev cleanup */
2108         if (is_mirroring_enc_context(kvm)) {
2109                 struct kvm *owner_kvm = sev->enc_context_owner;
2110
2111                 mutex_lock(&owner_kvm->lock);
2112                 list_del(&sev->mirror_entry);
2113                 mutex_unlock(&owner_kvm->lock);
2114                 kvm_put_kvm(owner_kvm);
2115                 return;
2116         }
2117
2118         /*
2119          * Ensure that all guest tagged cache entries are flushed before
2120          * releasing the pages back to the system for use. CLFLUSH will
2121          * not do this, so issue a WBINVD.
2122          */
2123         wbinvd_on_all_cpus();
2124
2125         /*
2126          * if userspace was terminated before unregistering the memory regions
2127          * then lets unpin all the registered memory.
2128          */
2129         if (!list_empty(head)) {
2130                 list_for_each_safe(pos, q, head) {
2131                         __unregister_enc_region_locked(kvm,
2132                                 list_entry(pos, struct enc_region, list));
2133                         cond_resched();
2134                 }
2135         }
2136
2137         sev_unbind_asid(kvm, sev->handle);
2138         sev_asid_free(sev);
2139 }
2140
2141 void __init sev_set_cpu_caps(void)
2142 {
2143         if (!sev_enabled)
2144                 kvm_cpu_cap_clear(X86_FEATURE_SEV);
2145         if (!sev_es_enabled)
2146                 kvm_cpu_cap_clear(X86_FEATURE_SEV_ES);
2147 }
2148
2149 void __init sev_hardware_setup(void)
2150 {
2151 #ifdef CONFIG_KVM_AMD_SEV
2152         unsigned int eax, ebx, ecx, edx, sev_asid_count, sev_es_asid_count;
2153         bool sev_es_supported = false;
2154         bool sev_supported = false;
2155
2156         if (!sev_enabled || !npt_enabled)
2157                 goto out;
2158
2159         /*
2160          * SEV must obviously be supported in hardware.  Sanity check that the
2161          * CPU supports decode assists, which is mandatory for SEV guests to
2162          * support instruction emulation.
2163          */
2164         if (!boot_cpu_has(X86_FEATURE_SEV) ||
2165             WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_DECODEASSISTS)))
2166                 goto out;
2167
2168         /* Retrieve SEV CPUID information */
2169         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
2170
2171         /* Set encryption bit location for SEV-ES guests */
2172         sev_enc_bit = ebx & 0x3f;
2173
2174         /* Maximum number of encrypted guests supported simultaneously */
2175         max_sev_asid = ecx;
2176         if (!max_sev_asid)
2177                 goto out;
2178
2179         /* Minimum ASID value that should be used for SEV guest */
2180         min_sev_asid = edx;
2181         sev_me_mask = 1UL << (ebx & 0x3f);
2182
2183         /*
2184          * Initialize SEV ASID bitmaps. Allocate space for ASID 0 in the bitmap,
2185          * even though it's never used, so that the bitmap is indexed by the
2186          * actual ASID.
2187          */
2188         nr_asids = max_sev_asid + 1;
2189         sev_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2190         if (!sev_asid_bitmap)
2191                 goto out;
2192
2193         sev_reclaim_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2194         if (!sev_reclaim_asid_bitmap) {
2195                 bitmap_free(sev_asid_bitmap);
2196                 sev_asid_bitmap = NULL;
2197                 goto out;
2198         }
2199
2200         sev_asid_count = max_sev_asid - min_sev_asid + 1;
2201         if (misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count))
2202                 goto out;
2203
2204         pr_info("SEV supported: %u ASIDs\n", sev_asid_count);
2205         sev_supported = true;
2206
2207         /* SEV-ES support requested? */
2208         if (!sev_es_enabled)
2209                 goto out;
2210
2211         /* Does the CPU support SEV-ES? */
2212         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
2213                 goto out;
2214
2215         /* Has the system been allocated ASIDs for SEV-ES? */
2216         if (min_sev_asid == 1)
2217                 goto out;
2218
2219         sev_es_asid_count = min_sev_asid - 1;
2220         if (misc_cg_set_capacity(MISC_CG_RES_SEV_ES, sev_es_asid_count))
2221                 goto out;
2222
2223         pr_info("SEV-ES supported: %u ASIDs\n", sev_es_asid_count);
2224         sev_es_supported = true;
2225
2226 out:
2227         sev_enabled = sev_supported;
2228         sev_es_enabled = sev_es_supported;
2229 #endif
2230 }
2231
2232 void sev_hardware_unsetup(void)
2233 {
2234         if (!sev_enabled)
2235                 return;
2236
2237         /* No need to take sev_bitmap_lock, all VMs have been destroyed. */
2238         sev_flush_asids(1, max_sev_asid);
2239
2240         bitmap_free(sev_asid_bitmap);
2241         bitmap_free(sev_reclaim_asid_bitmap);
2242
2243         misc_cg_set_capacity(MISC_CG_RES_SEV, 0);
2244         misc_cg_set_capacity(MISC_CG_RES_SEV_ES, 0);
2245 }
2246
2247 int sev_cpu_init(struct svm_cpu_data *sd)
2248 {
2249         if (!sev_enabled)
2250                 return 0;
2251
2252         sd->sev_vmcbs = kcalloc(nr_asids, sizeof(void *), GFP_KERNEL);
2253         if (!sd->sev_vmcbs)
2254                 return -ENOMEM;
2255
2256         return 0;
2257 }
2258
2259 /*
2260  * Pages used by hardware to hold guest encrypted state must be flushed before
2261  * returning them to the system.
2262  */
2263 static void sev_flush_encrypted_page(struct kvm_vcpu *vcpu, void *va)
2264 {
2265         int asid = to_kvm_svm(vcpu->kvm)->sev_info.asid;
2266
2267         /*
2268          * Note!  The address must be a kernel address, as regular page walk
2269          * checks are performed by VM_PAGE_FLUSH, i.e. operating on a user
2270          * address is non-deterministic and unsafe.  This function deliberately
2271          * takes a pointer to deter passing in a user address.
2272          */
2273         unsigned long addr = (unsigned long)va;
2274
2275         /*
2276          * If CPU enforced cache coherency for encrypted mappings of the
2277          * same physical page is supported, use CLFLUSHOPT instead. NOTE: cache
2278          * flush is still needed in order to work properly with DMA devices.
2279          */
2280         if (boot_cpu_has(X86_FEATURE_SME_COHERENT)) {
2281                 clflush_cache_range(va, PAGE_SIZE);
2282                 return;
2283         }
2284
2285         /*
2286          * VM Page Flush takes a host virtual address and a guest ASID.  Fall
2287          * back to WBINVD if this faults so as not to make any problems worse
2288          * by leaving stale encrypted data in the cache.
2289          */
2290         if (WARN_ON_ONCE(wrmsrl_safe(MSR_AMD64_VM_PAGE_FLUSH, addr | asid)))
2291                 goto do_wbinvd;
2292
2293         return;
2294
2295 do_wbinvd:
2296         wbinvd_on_all_cpus();
2297 }
2298
2299 void sev_guest_memory_reclaimed(struct kvm *kvm)
2300 {
2301         if (!sev_guest(kvm))
2302                 return;
2303
2304         wbinvd_on_all_cpus();
2305 }
2306
2307 void sev_free_vcpu(struct kvm_vcpu *vcpu)
2308 {
2309         struct vcpu_svm *svm;
2310
2311         if (!sev_es_guest(vcpu->kvm))
2312                 return;
2313
2314         svm = to_svm(vcpu);
2315
2316         if (vcpu->arch.guest_state_protected)
2317                 sev_flush_encrypted_page(vcpu, svm->sev_es.vmsa);
2318
2319         __free_page(virt_to_page(svm->sev_es.vmsa));
2320
2321         if (svm->sev_es.ghcb_sa_free)
2322                 kvfree(svm->sev_es.ghcb_sa);
2323 }
2324
2325 static void dump_ghcb(struct vcpu_svm *svm)
2326 {
2327         struct ghcb *ghcb = svm->sev_es.ghcb;
2328         unsigned int nbits;
2329
2330         /* Re-use the dump_invalid_vmcb module parameter */
2331         if (!dump_invalid_vmcb) {
2332                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
2333                 return;
2334         }
2335
2336         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
2337
2338         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
2339         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
2340                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
2341         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
2342                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
2343         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
2344                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
2345         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
2346                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
2347         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
2348 }
2349
2350 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
2351 {
2352         struct kvm_vcpu *vcpu = &svm->vcpu;
2353         struct ghcb *ghcb = svm->sev_es.ghcb;
2354
2355         /*
2356          * The GHCB protocol so far allows for the following data
2357          * to be returned:
2358          *   GPRs RAX, RBX, RCX, RDX
2359          *
2360          * Copy their values, even if they may not have been written during the
2361          * VM-Exit.  It's the guest's responsibility to not consume random data.
2362          */
2363         ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
2364         ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
2365         ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
2366         ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
2367 }
2368
2369 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
2370 {
2371         struct vmcb_control_area *control = &svm->vmcb->control;
2372         struct kvm_vcpu *vcpu = &svm->vcpu;
2373         struct ghcb *ghcb = svm->sev_es.ghcb;
2374         u64 exit_code;
2375
2376         /*
2377          * The GHCB protocol so far allows for the following data
2378          * to be supplied:
2379          *   GPRs RAX, RBX, RCX, RDX
2380          *   XCR0
2381          *   CPL
2382          *
2383          * VMMCALL allows the guest to provide extra registers. KVM also
2384          * expects RSI for hypercalls, so include that, too.
2385          *
2386          * Copy their values to the appropriate location if supplied.
2387          */
2388         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
2389
2390         vcpu->arch.regs[VCPU_REGS_RAX] = ghcb_get_rax_if_valid(ghcb);
2391         vcpu->arch.regs[VCPU_REGS_RBX] = ghcb_get_rbx_if_valid(ghcb);
2392         vcpu->arch.regs[VCPU_REGS_RCX] = ghcb_get_rcx_if_valid(ghcb);
2393         vcpu->arch.regs[VCPU_REGS_RDX] = ghcb_get_rdx_if_valid(ghcb);
2394         vcpu->arch.regs[VCPU_REGS_RSI] = ghcb_get_rsi_if_valid(ghcb);
2395
2396         svm->vmcb->save.cpl = ghcb_get_cpl_if_valid(ghcb);
2397
2398         if (ghcb_xcr0_is_valid(ghcb)) {
2399                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
2400                 kvm_update_cpuid_runtime(vcpu);
2401         }
2402
2403         /* Copy the GHCB exit information into the VMCB fields */
2404         exit_code = ghcb_get_sw_exit_code(ghcb);
2405         control->exit_code = lower_32_bits(exit_code);
2406         control->exit_code_hi = upper_32_bits(exit_code);
2407         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
2408         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
2409
2410         /* Clear the valid entries fields */
2411         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2412 }
2413
2414 static int sev_es_validate_vmgexit(struct vcpu_svm *svm)
2415 {
2416         struct kvm_vcpu *vcpu;
2417         struct ghcb *ghcb;
2418         u64 exit_code;
2419         u64 reason;
2420
2421         ghcb = svm->sev_es.ghcb;
2422
2423         /*
2424          * Retrieve the exit code now even though it may not be marked valid
2425          * as it could help with debugging.
2426          */
2427         exit_code = ghcb_get_sw_exit_code(ghcb);
2428
2429         /* Only GHCB Usage code 0 is supported */
2430         if (ghcb->ghcb_usage) {
2431                 reason = GHCB_ERR_INVALID_USAGE;
2432                 goto vmgexit_err;
2433         }
2434
2435         reason = GHCB_ERR_MISSING_INPUT;
2436
2437         if (!ghcb_sw_exit_code_is_valid(ghcb) ||
2438             !ghcb_sw_exit_info_1_is_valid(ghcb) ||
2439             !ghcb_sw_exit_info_2_is_valid(ghcb))
2440                 goto vmgexit_err;
2441
2442         switch (ghcb_get_sw_exit_code(ghcb)) {
2443         case SVM_EXIT_READ_DR7:
2444                 break;
2445         case SVM_EXIT_WRITE_DR7:
2446                 if (!ghcb_rax_is_valid(ghcb))
2447                         goto vmgexit_err;
2448                 break;
2449         case SVM_EXIT_RDTSC:
2450                 break;
2451         case SVM_EXIT_RDPMC:
2452                 if (!ghcb_rcx_is_valid(ghcb))
2453                         goto vmgexit_err;
2454                 break;
2455         case SVM_EXIT_CPUID:
2456                 if (!ghcb_rax_is_valid(ghcb) ||
2457                     !ghcb_rcx_is_valid(ghcb))
2458                         goto vmgexit_err;
2459                 if (ghcb_get_rax(ghcb) == 0xd)
2460                         if (!ghcb_xcr0_is_valid(ghcb))
2461                                 goto vmgexit_err;
2462                 break;
2463         case SVM_EXIT_INVD:
2464                 break;
2465         case SVM_EXIT_IOIO:
2466                 if (ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_STR_MASK) {
2467                         if (!ghcb_sw_scratch_is_valid(ghcb))
2468                                 goto vmgexit_err;
2469                 } else {
2470                         if (!(ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_TYPE_MASK))
2471                                 if (!ghcb_rax_is_valid(ghcb))
2472                                         goto vmgexit_err;
2473                 }
2474                 break;
2475         case SVM_EXIT_MSR:
2476                 if (!ghcb_rcx_is_valid(ghcb))
2477                         goto vmgexit_err;
2478                 if (ghcb_get_sw_exit_info_1(ghcb)) {
2479                         if (!ghcb_rax_is_valid(ghcb) ||
2480                             !ghcb_rdx_is_valid(ghcb))
2481                                 goto vmgexit_err;
2482                 }
2483                 break;
2484         case SVM_EXIT_VMMCALL:
2485                 if (!ghcb_rax_is_valid(ghcb) ||
2486                     !ghcb_cpl_is_valid(ghcb))
2487                         goto vmgexit_err;
2488                 break;
2489         case SVM_EXIT_RDTSCP:
2490                 break;
2491         case SVM_EXIT_WBINVD:
2492                 break;
2493         case SVM_EXIT_MONITOR:
2494                 if (!ghcb_rax_is_valid(ghcb) ||
2495                     !ghcb_rcx_is_valid(ghcb) ||
2496                     !ghcb_rdx_is_valid(ghcb))
2497                         goto vmgexit_err;
2498                 break;
2499         case SVM_EXIT_MWAIT:
2500                 if (!ghcb_rax_is_valid(ghcb) ||
2501                     !ghcb_rcx_is_valid(ghcb))
2502                         goto vmgexit_err;
2503                 break;
2504         case SVM_VMGEXIT_MMIO_READ:
2505         case SVM_VMGEXIT_MMIO_WRITE:
2506                 if (!ghcb_sw_scratch_is_valid(ghcb))
2507                         goto vmgexit_err;
2508                 break;
2509         case SVM_VMGEXIT_NMI_COMPLETE:
2510         case SVM_VMGEXIT_AP_HLT_LOOP:
2511         case SVM_VMGEXIT_AP_JUMP_TABLE:
2512         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2513                 break;
2514         default:
2515                 reason = GHCB_ERR_INVALID_EVENT;
2516                 goto vmgexit_err;
2517         }
2518
2519         return 0;
2520
2521 vmgexit_err:
2522         vcpu = &svm->vcpu;
2523
2524         if (reason == GHCB_ERR_INVALID_USAGE) {
2525                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
2526                             ghcb->ghcb_usage);
2527         } else if (reason == GHCB_ERR_INVALID_EVENT) {
2528                 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx is not valid\n",
2529                             exit_code);
2530         } else {
2531                 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx input is not valid\n",
2532                             exit_code);
2533                 dump_ghcb(svm);
2534         }
2535
2536         /* Clear the valid entries fields */
2537         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2538
2539         ghcb_set_sw_exit_info_1(ghcb, 2);
2540         ghcb_set_sw_exit_info_2(ghcb, reason);
2541
2542         /* Resume the guest to "return" the error code. */
2543         return 1;
2544 }
2545
2546 void sev_es_unmap_ghcb(struct vcpu_svm *svm)
2547 {
2548         if (!svm->sev_es.ghcb)
2549                 return;
2550
2551         if (svm->sev_es.ghcb_sa_free) {
2552                 /*
2553                  * The scratch area lives outside the GHCB, so there is a
2554                  * buffer that, depending on the operation performed, may
2555                  * need to be synced, then freed.
2556                  */
2557                 if (svm->sev_es.ghcb_sa_sync) {
2558                         kvm_write_guest(svm->vcpu.kvm,
2559                                         ghcb_get_sw_scratch(svm->sev_es.ghcb),
2560                                         svm->sev_es.ghcb_sa,
2561                                         svm->sev_es.ghcb_sa_len);
2562                         svm->sev_es.ghcb_sa_sync = false;
2563                 }
2564
2565                 kvfree(svm->sev_es.ghcb_sa);
2566                 svm->sev_es.ghcb_sa = NULL;
2567                 svm->sev_es.ghcb_sa_free = false;
2568         }
2569
2570         trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->sev_es.ghcb);
2571
2572         sev_es_sync_to_ghcb(svm);
2573
2574         kvm_vcpu_unmap(&svm->vcpu, &svm->sev_es.ghcb_map, true);
2575         svm->sev_es.ghcb = NULL;
2576 }
2577
2578 void pre_sev_run(struct vcpu_svm *svm, int cpu)
2579 {
2580         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2581         int asid = sev_get_asid(svm->vcpu.kvm);
2582
2583         /* Assign the asid allocated with this SEV guest */
2584         svm->asid = asid;
2585
2586         /*
2587          * Flush guest TLB:
2588          *
2589          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
2590          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
2591          */
2592         if (sd->sev_vmcbs[asid] == svm->vmcb &&
2593             svm->vcpu.arch.last_vmentry_cpu == cpu)
2594                 return;
2595
2596         sd->sev_vmcbs[asid] = svm->vmcb;
2597         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
2598         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
2599 }
2600
2601 #define GHCB_SCRATCH_AREA_LIMIT         (16ULL * PAGE_SIZE)
2602 static int setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
2603 {
2604         struct vmcb_control_area *control = &svm->vmcb->control;
2605         struct ghcb *ghcb = svm->sev_es.ghcb;
2606         u64 ghcb_scratch_beg, ghcb_scratch_end;
2607         u64 scratch_gpa_beg, scratch_gpa_end;
2608         void *scratch_va;
2609
2610         scratch_gpa_beg = ghcb_get_sw_scratch(ghcb);
2611         if (!scratch_gpa_beg) {
2612                 pr_err("vmgexit: scratch gpa not provided\n");
2613                 goto e_scratch;
2614         }
2615
2616         scratch_gpa_end = scratch_gpa_beg + len;
2617         if (scratch_gpa_end < scratch_gpa_beg) {
2618                 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
2619                        len, scratch_gpa_beg);
2620                 goto e_scratch;
2621         }
2622
2623         if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
2624                 /* Scratch area begins within GHCB */
2625                 ghcb_scratch_beg = control->ghcb_gpa +
2626                                    offsetof(struct ghcb, shared_buffer);
2627                 ghcb_scratch_end = control->ghcb_gpa +
2628                                    offsetof(struct ghcb, reserved_1);
2629
2630                 /*
2631                  * If the scratch area begins within the GHCB, it must be
2632                  * completely contained in the GHCB shared buffer area.
2633                  */
2634                 if (scratch_gpa_beg < ghcb_scratch_beg ||
2635                     scratch_gpa_end > ghcb_scratch_end) {
2636                         pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
2637                                scratch_gpa_beg, scratch_gpa_end);
2638                         goto e_scratch;
2639                 }
2640
2641                 scratch_va = (void *)svm->sev_es.ghcb;
2642                 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
2643         } else {
2644                 /*
2645                  * The guest memory must be read into a kernel buffer, so
2646                  * limit the size
2647                  */
2648                 if (len > GHCB_SCRATCH_AREA_LIMIT) {
2649                         pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
2650                                len, GHCB_SCRATCH_AREA_LIMIT);
2651                         goto e_scratch;
2652                 }
2653                 scratch_va = kvzalloc(len, GFP_KERNEL_ACCOUNT);
2654                 if (!scratch_va)
2655                         return -ENOMEM;
2656
2657                 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
2658                         /* Unable to copy scratch area from guest */
2659                         pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
2660
2661                         kvfree(scratch_va);
2662                         return -EFAULT;
2663                 }
2664
2665                 /*
2666                  * The scratch area is outside the GHCB. The operation will
2667                  * dictate whether the buffer needs to be synced before running
2668                  * the vCPU next time (i.e. a read was requested so the data
2669                  * must be written back to the guest memory).
2670                  */
2671                 svm->sev_es.ghcb_sa_sync = sync;
2672                 svm->sev_es.ghcb_sa_free = true;
2673         }
2674
2675         svm->sev_es.ghcb_sa = scratch_va;
2676         svm->sev_es.ghcb_sa_len = len;
2677
2678         return 0;
2679
2680 e_scratch:
2681         ghcb_set_sw_exit_info_1(ghcb, 2);
2682         ghcb_set_sw_exit_info_2(ghcb, GHCB_ERR_INVALID_SCRATCH_AREA);
2683
2684         return 1;
2685 }
2686
2687 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
2688                               unsigned int pos)
2689 {
2690         svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
2691         svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
2692 }
2693
2694 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
2695 {
2696         return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
2697 }
2698
2699 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
2700 {
2701         svm->vmcb->control.ghcb_gpa = value;
2702 }
2703
2704 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
2705 {
2706         struct vmcb_control_area *control = &svm->vmcb->control;
2707         struct kvm_vcpu *vcpu = &svm->vcpu;
2708         u64 ghcb_info;
2709         int ret = 1;
2710
2711         ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
2712
2713         trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
2714                                              control->ghcb_gpa);
2715
2716         switch (ghcb_info) {
2717         case GHCB_MSR_SEV_INFO_REQ:
2718                 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2719                                                     GHCB_VERSION_MIN,
2720                                                     sev_enc_bit));
2721                 break;
2722         case GHCB_MSR_CPUID_REQ: {
2723                 u64 cpuid_fn, cpuid_reg, cpuid_value;
2724
2725                 cpuid_fn = get_ghcb_msr_bits(svm,
2726                                              GHCB_MSR_CPUID_FUNC_MASK,
2727                                              GHCB_MSR_CPUID_FUNC_POS);
2728
2729                 /* Initialize the registers needed by the CPUID intercept */
2730                 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
2731                 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
2732
2733                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_CPUID);
2734                 if (!ret) {
2735                         /* Error, keep GHCB MSR value as-is */
2736                         break;
2737                 }
2738
2739                 cpuid_reg = get_ghcb_msr_bits(svm,
2740                                               GHCB_MSR_CPUID_REG_MASK,
2741                                               GHCB_MSR_CPUID_REG_POS);
2742                 if (cpuid_reg == 0)
2743                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
2744                 else if (cpuid_reg == 1)
2745                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
2746                 else if (cpuid_reg == 2)
2747                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
2748                 else
2749                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
2750
2751                 set_ghcb_msr_bits(svm, cpuid_value,
2752                                   GHCB_MSR_CPUID_VALUE_MASK,
2753                                   GHCB_MSR_CPUID_VALUE_POS);
2754
2755                 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
2756                                   GHCB_MSR_INFO_MASK,
2757                                   GHCB_MSR_INFO_POS);
2758                 break;
2759         }
2760         case GHCB_MSR_TERM_REQ: {
2761                 u64 reason_set, reason_code;
2762
2763                 reason_set = get_ghcb_msr_bits(svm,
2764                                                GHCB_MSR_TERM_REASON_SET_MASK,
2765                                                GHCB_MSR_TERM_REASON_SET_POS);
2766                 reason_code = get_ghcb_msr_bits(svm,
2767                                                 GHCB_MSR_TERM_REASON_MASK,
2768                                                 GHCB_MSR_TERM_REASON_POS);
2769                 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
2770                         reason_set, reason_code);
2771
2772                 ret = -EINVAL;
2773                 break;
2774         }
2775         default:
2776                 /* Error, keep GHCB MSR value as-is */
2777                 break;
2778         }
2779
2780         trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
2781                                             control->ghcb_gpa, ret);
2782
2783         return ret;
2784 }
2785
2786 int sev_handle_vmgexit(struct kvm_vcpu *vcpu)
2787 {
2788         struct vcpu_svm *svm = to_svm(vcpu);
2789         struct vmcb_control_area *control = &svm->vmcb->control;
2790         u64 ghcb_gpa, exit_code;
2791         struct ghcb *ghcb;
2792         int ret;
2793
2794         /* Validate the GHCB */
2795         ghcb_gpa = control->ghcb_gpa;
2796         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
2797                 return sev_handle_vmgexit_msr_protocol(svm);
2798
2799         if (!ghcb_gpa) {
2800                 vcpu_unimpl(vcpu, "vmgexit: GHCB gpa is not set\n");
2801
2802                 /* Without a GHCB, just return right back to the guest */
2803                 return 1;
2804         }
2805
2806         if (kvm_vcpu_map(vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->sev_es.ghcb_map)) {
2807                 /* Unable to map GHCB from guest */
2808                 vcpu_unimpl(vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
2809                             ghcb_gpa);
2810
2811                 /* Without a GHCB, just return right back to the guest */
2812                 return 1;
2813         }
2814
2815         svm->sev_es.ghcb = svm->sev_es.ghcb_map.hva;
2816         ghcb = svm->sev_es.ghcb_map.hva;
2817
2818         trace_kvm_vmgexit_enter(vcpu->vcpu_id, ghcb);
2819
2820         exit_code = ghcb_get_sw_exit_code(ghcb);
2821
2822         ret = sev_es_validate_vmgexit(svm);
2823         if (ret)
2824                 return ret;
2825
2826         sev_es_sync_from_ghcb(svm);
2827         ghcb_set_sw_exit_info_1(ghcb, 0);
2828         ghcb_set_sw_exit_info_2(ghcb, 0);
2829
2830         switch (exit_code) {
2831         case SVM_VMGEXIT_MMIO_READ:
2832                 ret = setup_vmgexit_scratch(svm, true, control->exit_info_2);
2833                 if (ret)
2834                         break;
2835
2836                 ret = kvm_sev_es_mmio_read(vcpu,
2837                                            control->exit_info_1,
2838                                            control->exit_info_2,
2839                                            svm->sev_es.ghcb_sa);
2840                 break;
2841         case SVM_VMGEXIT_MMIO_WRITE:
2842                 ret = setup_vmgexit_scratch(svm, false, control->exit_info_2);
2843                 if (ret)
2844                         break;
2845
2846                 ret = kvm_sev_es_mmio_write(vcpu,
2847                                             control->exit_info_1,
2848                                             control->exit_info_2,
2849                                             svm->sev_es.ghcb_sa);
2850                 break;
2851         case SVM_VMGEXIT_NMI_COMPLETE:
2852                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_IRET);
2853                 break;
2854         case SVM_VMGEXIT_AP_HLT_LOOP:
2855                 ret = kvm_emulate_ap_reset_hold(vcpu);
2856                 break;
2857         case SVM_VMGEXIT_AP_JUMP_TABLE: {
2858                 struct kvm_sev_info *sev = &to_kvm_svm(vcpu->kvm)->sev_info;
2859
2860                 switch (control->exit_info_1) {
2861                 case 0:
2862                         /* Set AP jump table address */
2863                         sev->ap_jump_table = control->exit_info_2;
2864                         break;
2865                 case 1:
2866                         /* Get AP jump table address */
2867                         ghcb_set_sw_exit_info_2(ghcb, sev->ap_jump_table);
2868                         break;
2869                 default:
2870                         pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
2871                                control->exit_info_1);
2872                         ghcb_set_sw_exit_info_1(ghcb, 2);
2873                         ghcb_set_sw_exit_info_2(ghcb, GHCB_ERR_INVALID_INPUT);
2874                 }
2875
2876                 ret = 1;
2877                 break;
2878         }
2879         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2880                 vcpu_unimpl(vcpu,
2881                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
2882                             control->exit_info_1, control->exit_info_2);
2883                 ret = -EINVAL;
2884                 break;
2885         default:
2886                 ret = svm_invoke_exit_handler(vcpu, exit_code);
2887         }
2888
2889         return ret;
2890 }
2891
2892 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
2893 {
2894         int count;
2895         int bytes;
2896         int r;
2897
2898         if (svm->vmcb->control.exit_info_2 > INT_MAX)
2899                 return -EINVAL;
2900
2901         count = svm->vmcb->control.exit_info_2;
2902         if (unlikely(check_mul_overflow(count, size, &bytes)))
2903                 return -EINVAL;
2904
2905         r = setup_vmgexit_scratch(svm, in, bytes);
2906         if (r)
2907                 return r;
2908
2909         return kvm_sev_es_string_io(&svm->vcpu, size, port, svm->sev_es.ghcb_sa,
2910                                     count, in);
2911 }
2912
2913 void sev_es_init_vmcb(struct vcpu_svm *svm)
2914 {
2915         struct kvm_vcpu *vcpu = &svm->vcpu;
2916
2917         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
2918         svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
2919
2920         /*
2921          * An SEV-ES guest requires a VMSA area that is a separate from the
2922          * VMCB page. Do not include the encryption mask on the VMSA physical
2923          * address since hardware will access it using the guest key.
2924          */
2925         svm->vmcb->control.vmsa_pa = __pa(svm->sev_es.vmsa);
2926
2927         /* Can't intercept CR register access, HV can't modify CR registers */
2928         svm_clr_intercept(svm, INTERCEPT_CR0_READ);
2929         svm_clr_intercept(svm, INTERCEPT_CR4_READ);
2930         svm_clr_intercept(svm, INTERCEPT_CR8_READ);
2931         svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
2932         svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
2933         svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
2934
2935         svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
2936
2937         /* Track EFER/CR register changes */
2938         svm_set_intercept(svm, TRAP_EFER_WRITE);
2939         svm_set_intercept(svm, TRAP_CR0_WRITE);
2940         svm_set_intercept(svm, TRAP_CR4_WRITE);
2941         svm_set_intercept(svm, TRAP_CR8_WRITE);
2942
2943         /* No support for enable_vmware_backdoor */
2944         clr_exception_intercept(svm, GP_VECTOR);
2945
2946         /* Can't intercept XSETBV, HV can't modify XCR0 directly */
2947         svm_clr_intercept(svm, INTERCEPT_XSETBV);
2948
2949         /* Clear intercepts on selected MSRs */
2950         set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
2951         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
2952         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
2953         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
2954         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
2955         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
2956 }
2957
2958 void sev_es_vcpu_reset(struct vcpu_svm *svm)
2959 {
2960         /*
2961          * Set the GHCB MSR value as per the GHCB specification when emulating
2962          * vCPU RESET for an SEV-ES guest.
2963          */
2964         set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2965                                             GHCB_VERSION_MIN,
2966                                             sev_enc_bit));
2967 }
2968
2969 void sev_es_prepare_switch_to_guest(struct vmcb_save_area *hostsa)
2970 {
2971         /*
2972          * As an SEV-ES guest, hardware will restore the host state on VMEXIT,
2973          * of which one step is to perform a VMLOAD.  KVM performs the
2974          * corresponding VMSAVE in svm_prepare_guest_switch for both
2975          * traditional and SEV-ES guests.
2976          */
2977
2978         /* XCR0 is restored on VMEXIT, save the current host value */
2979         hostsa->xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
2980
2981         /* PKRU is restored on VMEXIT, save the current host value */
2982         hostsa->pkru = read_pkru();
2983
2984         /* MSR_IA32_XSS is restored on VMEXIT, save the currnet host value */
2985         hostsa->xss = host_xss;
2986 }
2987
2988 void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
2989 {
2990         struct vcpu_svm *svm = to_svm(vcpu);
2991
2992         /* First SIPI: Use the values as initially set by the VMM */
2993         if (!svm->sev_es.received_first_sipi) {
2994                 svm->sev_es.received_first_sipi = true;
2995                 return;
2996         }
2997
2998         /*
2999          * Subsequent SIPI: Return from an AP Reset Hold VMGEXIT, where
3000          * the guest will set the CS and RIP. Set SW_EXIT_INFO_2 to a
3001          * non-zero value.
3002          */
3003         if (!svm->sev_es.ghcb)
3004                 return;
3005
3006         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, 1);
3007 }